Download Part 6 -- Convergence of Random Variables Variables and Limit Theorems

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Markov and Chebyshev Inequalities
Recall that a random variable X is called nonnegative if P (X
STAT/MTHE 353:
6 – Convergence of Random Variables
and Limit Theorems
0) = 1.
Theorem 1 (Markov’s inequality)
Let X be a nonnegative random variable with mean E(X). Then for any
t>0
E(X)
P (X t) 
t
T. Linder
Proof: Assume X is continuous with pdf f . Then f (x) = 0 if x < 0, so
Z 1
Z 1
Z 1
E(X) =
xf (x) dx
xf (x) dx t
f (x) dx
Queen’s University
0
=
t
tP (X
t
t)
Winter 2012
⇤
If X is discrete, replace the integrals with sums. . .
STAT/MTHE 353: 6 – Convergence and Limit Theorems
1 / 34
Let X be a nonnegative random variable with finite variance Var(X).
Then for any t > 0
E(X)|
t 
E(X)|
t)
=
=
P (|X
Let X be a random variable with MGF MX (t). Then for any a 2 R
Var(X)
t2
E(X)|2
Var(X)
t2
Example: Chebyshev with t = k
STAT/MTHE 353: 6 – Convergence and Limit Theorems
t2 ) 
2 / 34
Theorem 3 (Cherno↵’s bound)
a)  min e
P (X
Proof: Apply Markov’s inequality to the nonnegative random variable
Y = |X E(X)|2
P (|X
STAT/MTHE 353: 6 – Convergence and Limit Theorems
The following result often gives a much sharper bound if the MGF of X
is finite in some interval around zero.
Theorem 2 (Chebyshev’s inequality)
P |X
Example: Suppose X is nonnegative and P (X > 10) = 1/5; show that
E(X) 2. Also, Markov’s inequality for |X|.
E |X
t>0
at
MX (t)
Proof: Fix t > 0. Then we have
E(X)|2
P (X
a)
=
t2
P (tX
ta) = P etX
eta
tX

⇤
=
p
Var(X) . . .
E e
(Markov’s inequality)
eta
e ta MX (t)
Since this holds for all t > 0, it must hold for the t minimizing the upper
bound.
⇤
3 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
4 / 34
Convergence of Random Variables
A probability space is a triple (⌦, A, P ), where ⌦ is a sample space, A is
a collection of subsets of ⌦ called events, and P is a probability measure
on A. In particular, the set of events A satisfies
Example: Suppose X ⇠ N (0, 1). Apply Cherno↵’s bound to upper
bound P (X a) for a > 0 and compare with the bounds obtained from
Chebyshev’s and Markov’s inequalities.
1) ⌦ is an event
2) If A ⇢ ⌦ is an event, then Ac is also an event
S1
3) If A1 , A2 , A3 , . . . are events, then so is n=1 An .
P is a function from the collection of events A to [0, 1] which satisfies
the axioms of probability:
1) P (A)
0 for all events A 2 A.
2) P (⌦) = 1.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
5 / 34
3) If A1 , A2 , A3 , . . . are mutually exclusive events (i.e., Ai \ Aj = ; for
S1
P1
all i 6= j), then P
i=1 Ai =
i=1 P (Ai ).
STAT/MTHE 353: 6 – Convergence and Limit Theorems
6 / 34
Definitions (Modes of convergence) Let {Xn } = X1 , X2 , X3 , . . . be a
sequence of random variables defined in a probability space (⌦, A, P ).
(i) We say that {Xn } converges to a random variable X almost surely
a.s.
(notation: Xn ! X) if
Recall that a random variable X is a function X : ⌦ ! R that maps
any point ! in the sample space ⌦ to a real number X(!).
P ({! : Xn (!) ! X(!)}) = 1
A random variable X must satisfy the following: any subset A of ⌦
in the form
A = {! 2 ⌦ : X(!) 2 B}
(i) We say that {Xn } converges to a random variable X in probability
P
(notation: Xn ! X) if for any ✏ > 0 we have
P |Xn
for any “reasonable” B ⇢ R is an event. For example B can be any
set obtained by a countable union and intersection of intervals.
X| > ✏ ! 0
as
n!1
(iii) If r > 0 we say that {Xn } converges to a random variable X in rth
r.m.
mean (notation: Xn ! X) if
Recall that a sequence of real numbers {xn }1
n=1 is said to converge
to a limit x 2 R (notation: xn ! x) if for any ✏ > 0 there exists N
such that |xn x| < ✏ for all n N .
E |Xn
X|r ! 0
as
n!1
(iv) Let Fn and F denote the cdfs of Xn and X, respectively. We say
that {Xn } converges to a random variable X in distribution
d
(notation: Xn ! X) if
Fn (x) ! F (x)
as
n!1
for any x such that F is continuous at x.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
7 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
8 / 34
Remarks:
Theorem 4
The following implications hold:
A very important case of convergence in rth mean is when r = 2. In
this case E |Xn X|2 ! 0 as n ! 1 and we say that {Xn }
converges in mean square to X.
Xn
a.s.
!X
Almost sure convergence is often called convergence with
probability 1.
"*
Xn
4<
Almost sure convergence, convergence in rth mean, and
convergence in probability all state that Xn is eventually close to X
(in di↵erent senses) as n increases.
Xn
In contrast, convergence in distribution is only a statement about
the closeness of the distribution of Xn to that of X for large n.
P
+3 Xn d! X
!X
r.m.
!X
a.s.
Remark: We will show that in general, Xn ! X does not imply that
r.m.
r.m.
a.s.
Xn ! X, and also that Xn ! X does not imply Xn ! X.
Example: Sequence {Xn } such that Fn (x) = F (x) for all x 2 R and
n = 1, 2, . . ., but P (|Xn X| > 1/2) = 1 for all n . . .
STAT/MTHE 353: 6 – Convergence and Limit Theorems
9 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
10 / 34
Theorem 6
Almost sure convergence implies convergence in probability; i.e., if
a.s.
P
Xn ! X, then Xn ! X.
Theorem 5
Convergence in rth mean implies convergence in probability; i.e., if
r.m.
P
Xn ! X, then Xn ! X.
r.m.
Proof: Assume Xn ! X for some r > 0; i.e., E(|Xn
n ! 1. Then for any ✏ > 0,
P (|Xn
X| > ✏)
=

X|r ) ! 0 as
An (✏) = ! : |Xn (!)
X(!)| > ✏
we want to show that P An (✏) ! 0 for any ✏ > 0.
P (|Xn X|r > ✏r )
E(|Xn X|r )
! 0 as n ! 1
✏r
where the inequality follows from Markov’s inequality applied to the
nonnegative random variable |Xn X|r .
STAT/MTHE 353: 6 – Convergence and Limit Theorems
a.s.
Proof: Assume Xn ! X. We want to show that for any ✏ > 0 we have
P (|Xn X| > ✏) ! 0 as n ! 1. Thus, defining the event
Define
A(✏) = ! : |Xn (!)
⇤
X(!)| > ✏ for infinitely many n
Then for any ! 2 A(✏) we have that Xn (!) 6! X(!). Since Xn
we must have that
P A(✏) = 0
11 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
a.s.
!X
12 / 34
Proof cont’d: Now define the event
Bn (✏) = ! : |Xm (!)
X(!)| > ✏ for some m
n
Proof cont’d: We clearly have
Notice that B1 (✏) B2 (✏) · · · Bn 1 (✏) Bn (✏)
Bn (✏) is a decreasing sequence of events, satisfying
A(✏) =
1
\
· · · ; i.e.,
An (✏) ⇢ Bn (✏)
so P An (✏)  P Bn (✏) , so P An (✏) ! 0 as n ! 1. We obtain
Bn (✏)
n=1
lim P |Xn
n!1
Therefore by the continuity of probability
P A(✏) = P
✓\
1
Bn (✏)
n=1
◆
for all n
X| > ✏ = 0
⇤
for any ✏ > 0 as claimed.
= lim P Bn (✏)
n!1
But since P A(✏) = 0, we obtain that lim P Bn (✏) = 0.
n!1
STAT/MTHE 353: 6 – Convergence and Limit Theorems
13 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
14 / 34
Lemma 7 (Sufficient condition for a.s. convergence)
Suppose
P
a.s.
n=1
Example: (Xn ! X does not imply Xn ! X.) Let X1 , X2 , . . . be
independent random variables with distribution
P (Xn = 0) = 1
1
,
n
r.m.
! X does not imply Xn
X| > ✏) < 1 for all ✏ > 0. Then Xn
A(✏) = ! : |Xn (!)
P
! X, but
a.s.
! X.
X(!)| > ✏ for infinitely many n
Then
A(✏)c = ! : there exists N such that |Xn (!) X(!)|  ✏ for all n
Solution: . . .
Example: (Xn
example. . .
P (|Xn
Proof: Recall the event
1
P (Xn = 1) =
n
and show that there is a random variable X such that Xn
P ({! : Xn (!) ! X(!)}) = 0.
1
P
For ✏ = 1/k we obtain a decreasing sequence of events A(1/k)c
Let
1
\
C=
A(1/k)c
a.s.
! X.) Use the previous
N
1
.
k=1
k=1
and notice that if ! 2 C, then Xn (!) ! X(!) as n ! 1. Thus if
a.s.
P (C) = 1, then Xn ! X.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
15 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
16 / 34
Proof cont’d: Note that
Proof cont’d: But from the continuity of probability
P (C) = P
✓\
1
A(1/k)c
k=1
◆
Bn (✏) =
= lim P A(1/k)c
Bn (✏) = ! : |Xm (!)
and that the condition
implies
P1
n!1
P Bn (✏)
=
n!1

Thus if we can show that lim P (Bn (✏)) = 0 for all ✏ > 0, then
n!1
! X.
P1
n=1
P
✓ [
1
Am (✏)
1
X
m=n
◆
(union bound)
P Am (✏) ! 0 as n ! 1
n!1
17 / 34
X| > ✏) < 1
P An (✏) = 0
so lim P (Bn (✏)) = 0 for all ✏ > 0. Thus Xn
STAT/MTHE 353: 6 – Convergence and Limit Theorems
P (|Xn
m=n
m=n
P (A(✏)) = lim P (Bn (✏))
a.s.
1
X
We obtain
n
We have seen in the proof of Theorem 6 that
Xn
P An (✏) =
n=1
lim
X(!)| > ✏ and
X(!)| > ✏ for some m
Am (✏)
m=n
k!1
so if P A(1/k)c = 1 for all k, then P (C) = 1 and we obtain
a.s.
a.s.
Xn ! X. Thus P A(✏) = 0 for all ✏ > 0 implies Xn ! X.
As before, let An (✏) = ! : |Xn (!)
1
[
a.s.
! X.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
⇤
18 / 34
Theorem 8
a.s.
Convergence in probability implies convergence in distribution; i.e., if
P
d
Xn ! X, then Xn ! X.
r.m.
Example: (Xn ! X does not imply Xn ! X.) Let X1 , X2 , . . . be
random variables with marginal distributions given by
P (Xn = n3 ) =
1
,
n2
P (Xn = 0) = 1
Show that there is a random variable X such that Xn
r.m.
Xn 6 ! X if r 2/3.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
Proof: Let Fn and F be the cdfs of Xn and X, respectively and let
x 2 R be such that F is continuous at x. We want to show that
P
Xn ! X implies Fn (x) ! F (x) as n ! 1.
1
n2
For any given ✏ > 0 we have
a.s.
! X, but
Fn (x)
19 / 34
=
P (Xn  x)
=
P Xn  x, X  x + ✏ + P Xn  x, X > x + ✏

P X  x + ✏ + P |Xn
=
F (x + ✏) + P |Xn
STAT/MTHE 353: 6 – Convergence and Limit Theorems
X| > ✏
X| > ✏
20 / 34
If X is a constant random variable, then the converse also holds:
Proof cont’d: Similarly,
Theorem 9
F (x
✏)
=
P (X  x
✏)
=
P Xx
✏, Xn  x + P X  x

P Xn  x + P |Xn
=
Fn (x) + P |Xn
d
Let c 2 R. If Xn ! c, then Xn
✏, Xn > x
X| > ✏
P |Xn
✏)
X| > ✏
X| > ✏  Fn (x)  F (x + ✏) + P |Xn
X| > ✏
Then F is continuous at all x 6= c, so Fn (x) ! F (x) as n ! 1 for all
x 6= c. For any ✏ > 0
P
Since Xn ! X, we have P |Xn X| > ✏ ! 0 as n ! 1. Choosing
N (✏) large enough so that P |Xn X| > ✏ < ✏ for all n N (✏), we
obtain
F (x ✏) ✏  Fn (x)  F (x + ✏) + ✏
P |Xn
for all n N (✏). Since F is continuous at x, letting ✏ ! 0 we obtain
limn!1 Fn (x) = F (x).
⇤
STAT/MTHE 353: 6 – Convergence and Limit Theorems
! c.
Proof: For X with P (X = c) let Fn and F be as before and note that
8
<0 if x < c
F (x) =
:1 if x c
We obtain
F (x
P
21 / 34
c| > ✏
=
P (Xn < c
✏) + P (Xn > c + ✏)

P (Xn  c
✏) + P (Xn > c + ✏)
=
Fn (c
✏) + 1
Fn (c + ✏)
Since Fn (c ✏) ! 0 and Fn (c + ✏) ! 1 as n ! 1, we obtain
P |Xn c| > ✏ ! 0 as n ! 1.
⇤
STAT/MTHE 353: 6 – Convergence and Limit Theorems
22 / 34
Laws of Large Numbers
Proof: Since the Xi are independent,
Let X1 , X2 , . . . be a sequence of independent and identically distributed
(i.i.d.) random variables with finite mean µ = E(Xi ) and variance
2
= Var(Xi ). The sample mean X̄n is defined by
Var(X̄n ) = Var
n
X̄n =
1X
Xi
n i=1
P |X̄n
Thus P |X̄n
P
! µ; i.e., for all ✏ > 0
lim P |X̄n
n!1
STAT/MTHE 353: 6 – Convergence and Limit Theorems
n
1X
Xi
n i=1
◆
=
n
2
1 X
Var(X
)
=
i
n2 i=1
n
Since E(X̄n ) = µ, Chebyshev’s inequality implies for any ✏ > 0
Theorem 10 (Weak law of large numbers)
We have X̄n
✓
µ| > ✏ 
Var(X̄n )
= 2
✏2
✏ n
µ| > ✏ ! 0 as n ! 1 for any ✏ > 0.
⇤
µ| > ✏ = 0
23 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
24 / 34
Theorem 11 (Strong law of large numbers)
Proof cont’d: Next suppose that Xi 0 for all i. Then
Sn (!) = X1 (!) + · · · + Xn (!) is a nondecreasing sequence. For any n
there is a unique in such that i2n  n < (in + 1)2 . Thus
If X1 , X2 , . . . is an i.i.d. sequence with finite mean µ = E(Xi ) and
a.s.
variance Var(Xi ) = 2 , then X̄n ! µ; i.e.,
P
✓
n
1X
Xi = µ
n!1 n
i=1
lim
◆
Si2n  Sn  S(in +1)2
=1
1
P
n=1
P |X̄n2
µ| > ✏ < 1 and Lemma 7 gives X̄n2
A1 =
0, Xi
n
!:
1X +
X (!) ! µ+ ,
n i=1 i
A2 =
⇢
n
n
1X
Xi (!)
n!1 n
i=1
We conclude that
1
n
=
=
Pn
i=1
1X +
Xi (!)
n!1 n
i=1
lim
µ+
Xi
STAT/MTHE 353: 6 – Convergence and Limit Theorems
µ =µ
a.s.
! µ.
X̄i2n  X̄n 
✓
in + 1
in
◆2
X̄(in +1)2
STAT/MTHE 353: 6 – Convergence and Limit Theorems
26 / 34
Example: Simple random walk . . .
0. Letting
Example: (Single server queue) Customers arrive one-by-one at a service
station.
1X
X (!) ! µ
n i=1 i
The time between the arrival of the (i 1)th and ith customer is Yi ,
and Y1 , Y2 , . . . are i.i.d. nonnegative random variables with finite
mean E(Y1 ) and finite variance.
we know that P (A1 ) = P (A2 ) = 1. Thus P (A1 \ A2 ) = 1. But for all
! 2 A1 \ A2 we have
lim
◆2
Remark: The condition Var(Xi ) < 1 is not needed. The strong law of
large numbers (SLLN) holds for any i.i.d. sequence X1 , X2 , . . . with finite
mean µ = E(X1 ).
0. Define
n
!:
in
in + 1
X̄n (!) ! µ for all ! 2 A
! µ.
Xi = max( Xi , 0)
and note that Xi = Xi+ Xi and Xi+
µ+ = E(Xi+ ), µ = E(Xi ) and
✓
Letting A = {! : X̄n2 (!) ! µ}, we know that P (A) = 1. Since
in /(in + 1) ! 1 and (in + 1)/in ! 1 as n ! 1, and for all ! 2 A,
X̄i2n (!) ! µ and X̄(in +1)2 (!) ! µ as n ! 1, we obtain
25 / 34
Xi+ = max(Xi , 0),
⇢
i.e.
a.s.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
Proof cont’d: Now we remove the restriction Xi
Si2n
S(in +1)2
Sn


(in + 1)2
n
i2n
This is equivalent to
✓
◆2
✓
◆2
Si2n
S(in +1)2
in
Sn
in + 1


in + 1
i2n
n
in
(in + 1)2
Proof: First we show that the subsequence X̄12 , X̄22 , X̄32 , . . . converges
a.s. to µ.
Pn
Let Sn = i=1 Xi . Then E(Sn2 ) = n2 µ and Var(Sn2 ) = n2 2 . For any
✏ > 0 we have by Chebyshev’s inequality
◆
✓
1
P |X̄n2 µ| > ✏ = P
Sn2 µ > ✏ = P |Sn2 n2 µ| > n2 ✏
n2
2
Var(Sn2 )

=
n4 ✏ 2
n2 ✏ 2
Thus
=)
The time needed to service the ith customer is Ui , and U1 , U2 , . . .
are i.i.d. nonnegative random variables with finite mean E(U1 ) and
finite variance.
n
1X
Xi (!)
n!1 n
i=1
lim
Show that if E(U1 ) < E(Y1 ), then (after the first customer arrives) the
queue will eventually become empty with probability 1.
⇤
27 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
28 / 34
Central Limit Theorem
The De Moivre-Laplace theorem is a special case of the following general
result:
If X1 , X2 , . . . are Bernoulli(p) random variables, then
Sn = X1 + · · · + Xn is a Binomial(n, p) random variable with mean
E(Sn ) = np and variance Var(Sn ) = np(1 p).
Recall the De Moivre-Laplace theorem:
✓
◆
Sn np
lim P p
x =
n!1
np(1 p)
where (x) =
variable X.
Since
Rx
1
2
p1 e t /2
2⇡
Theorem 12 (Central Limit Theorem)
Let X1 , X2 , . . . be i.i.d. random variables with mean µ and finite
variance 2 . Then for Sn = X1 + · · · + Xn we have
Sn
(x)
nµ d
!X
n
where X ⇠ N (0, 1).
dt is the cdf of a N (0, 1) random
p
Remark: Note that Sn pnµ
= n X̄n µ . The SLLN implies that with
n
probability 1, X̄n µ ! 0 as n ! 1 The central limit theorem (CLT)
tells us something about the speed at which X̄n µ converges to zero.
(x) is continuous at every x, the above is equivalent to
Sn
where µ = E(X1 ) = p and
STAT/MTHE 353: 6 – Convergence and Limit Theorems
p
nµ d
!X
n
p
= Var(X1 ) = p(1
p
p).
29 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
30 / 34
Proof of CLT: Let Yn = Xn µ. Then E(Yn ) = 0 and Var(Yn ) =
Note that
Pn
Sn nµ
i=1 Yi
p
p
=
= Zn
n
n
To prove the CLT we will assume that each Xi has MGF MXi (t) that is
defined in an open interval around zero. The key to the proof is the
following result which we won’t prove:
2
.
Letting MY (t) denote the (common) moment generating function of the
Yi , we have by the independence of Y1 , Y2 , . . .
✓
◆n
t
p
MZn (t) = MY
n
Theorem 13 (Levy continuity theorem for MGF)
Assume Z1 , Z2 , . . . are random variables such that the MGF MZn (t) is
defined for all t 2 ( a, a) for some a > 0 and all n = 1, 2, . . . Suppose X
is a random variable with MGF MX (t) defined for t 2 ( a, a). If
If X ⇠ N (0, 1), then MX (t) = et
2
MZn (t) ! et /2 , or equivalently
lim MZn (t) = MX (t) for all t 2 ( a, a)
n!1
2
/2
. Thus if we show that
d
then Zn ! X.
lim ln MZn (t) =
n!1
t2
2
then Levy’s continuity theorem implies the CLT.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
31 / 34
STAT/MTHE 353: 6 – Convergence and Limit Theorems
32 / 34
Proof cont’d: Let h =
t
p . Then
n
lim ln MZn (t)
n!1
=
=
=
Proof cont’d: Apply l’Hospital’s rule again:
◆n ◆
✓
✓
t
p
lim ln MY
n!1
n
✓
✓
◆◆
t
p
lim n ln MY
n!1
n
t2
ln MY (h)
lim
2 h!0
h2
lim ln MZn (t)
n!1
=
=
=
It can be shown that MY (h) and all its derivatives are continuous at
h = 0. Since MY (0) = E(e0·Y ) = 1, the above limit is indeterminate.
Applying l’Hospital’s rule, we consider the limit
t2
lim
2 h!0
t2
ln MY (h)
h2
t
MY0 (h)
lim
2 h!0 2hM (h)
Y
t2
MY00 (h)
lim
2 h!0 2M (h) + 2hM 0 (h)
Y
Y
lim
2 h!0
2
t2
2
·
2
MY00 (0)
t2
= 2·
2MY (0)
2
2
=
MY0 (h)/MY (h)
t2
MY0 (h)
= 2 lim
h!0 2hMY (h)
2h
Thus MZn (t) ! et
2
/2
33 / 34
t
2
and therefore Zn =
X ⇠ N (0, 1).
Again, since M 0 (0) = E(Y ) = 0, the limit is indeterminate.
STAT/MTHE 353: 6 – Convergence and Limit Theorems
=
STAT/MTHE 353: 6 – Convergence and Limit Theorems
Sn
p
nµ d
! X, where
n
⇤
34 / 34